v0.20.1

Two fixes for a production outage first seen in 0.20.0: every session row disappearing from the left pane after a 5xx response, and Settings → Git hosting never finishing its first load.

Fixed

• Every session row disappeared after any 5xx response and did not come back on its own. The api() helper returns an error object rather than throwing, so the sessions store treated the response as a valid empty list and the rail went blank. The polling that would have recovered it stopped while a push connection was healthy, and reconnecting only re-fetched the current user — not the session list. A browser reload now restores it immediately.
• Settings → Git hosting never finished loading, and could take the whole /connections request to a 60-second load-balancer timeout. The agent called claude auth status synchronously, with no timeout, on every request; in practice this CLI took 21–28 seconds because it contested a file lock with other Claude sessions sharing the same config directory. The fix runs the probe asynchronously: if it does not answer in time the previous known result is returned, and the fresh answer arrives at the next poll.

Upgrade notes

No action needed on the ECS deployment — the Control Plane image update is enough for the Console fix.

All workspace users must restart their workspaces to receive the agent fix. Workspaces that were running during or before the upgrade are still on the previous agent binary; the Console will show a “restart required” badge until they do.